A centrifugal pump variable working condition multi-state intelligent diagnosis method and system based on motor q-axis current time-frequency feature fusion

Abstract

The application provides a centrifugal pump variable working condition multi-state intelligent diagnosis method and system based on motor q-axis current time-frequency feature fusion, comprising an offline model training stage and an online intelligent diagnosis stage. In the offline stage, three-phase stator currents and synchronous electric angles under multiple working conditions are collected, q-axis currents are extracted through Clarke / Park transformation, time-frequency dual-domain features are constructed through multi-layer discrete wavelet decomposition and reconstruction and Fourier transformation after pretreatment, a multi-task deep diagnosis model containing double-branch CNN, dynamic soft mask, bidirectional long short-term memory network and multi-head attention is trained and parameters are solidified. In the online stage, currents and synchronous electric angles are collected in real time, online time-frequency dual-domain features are constructed and input into the trained model, solidified parameters are called to synchronously output flow working conditions, cavitation stages and port ring wear states. Non-invasive multi-state joint diagnosis is realized only by using motor power supply loop signals.

Description

Technical Field

[0001] This invention relates to the fields of fluid machinery condition monitoring, motor current characteristic analysis and intelligent fault diagnosis, and in particular to a multi-state intelligent diagnosis method and system for centrifugal pumps under varying operating conditions based on the fusion of time-frequency characteristics of motor q-axis current. Background Technology

[0002] Centrifugal pumps are widely used in fluid transport applications such as petrochemicals, water conservancy projects, energy equipment, agricultural irrigation, and process industries. During long-term operation, centrifugal pumps are affected by factors such as flow rate deviations from the design point, inlet pressure changes, fluid pulsation, and seal clearance degradation, leading to cavitation, wear of the sealing rings, and performance decline. There is also a coupling relationship between flow rate changes, cavitation evolution, and sealing ring wear. For example, changes in flow rate cause changes in the pressure distribution within the pump, clearance leakage caused by sealing ring wear alters the pump's head and load, and cavitation impacts can accelerate damage to flow-through components. Therefore, testing under only a single condition is insufficient to comprehensively assess the operational health of a centrifugal pump.

[0003] Current centrifugal pump condition monitoring typically uses pressure, flow, vibration, or acoustic signals. While these signals can reflect the hydraulic state within the pump, they require dedicated sensors installed in the pump body, bearing housing, or pipelines, and are easily affected by installation location, structural transmission path, fluid medium, on-site noise, and maintenance conditions. In scenarios involving submersible pumps, high temperatures, high pressures, or where pipeline modifications are inconvenient, the long-term reliability and installation cost of external sensors become even more critical.

[0004] The drive motor and centrifugal pump are mechanically connected via a shaft system. Steady-state deviations and transient fluctuations in the hydraulic load torque on the pump side will cause corresponding changes in the motor's electromagnetic torque and stator current. Monitoring the mechanical load using the existing current signal in the power supply circuit does not require intrusion into the pump body or flow channel, and is characterized by convenient installation, low cost, and suitability for online applications. However, when directly using the three-phase current amplitude or a single spectrum characteristic, abnormal features are easily masked by the power supply fundamental frequency, inverter switching harmonics, and random noise; at the same time, flow changes, cavitation impacts, and wear of the flow rings manifest as disturbances on different time and frequency scales, which are difficult to account for simultaneously in a single time-domain or frequency-domain analysis.

[0005] In a synchronous rotating coordinate system, the q-axis current is closely related to the motor's electromagnetic torque and can more effectively characterize load torque variations than the original phase current. However, existing technologies typically fail to address the following issues simultaneously: how to separate multi-scale weak features from the q-axis current; how to deeply fuse time-domain transient information and frequency-domain modulation information; how to suppress irrelevant channels under inverter noise and flow variation conditions; and how to utilize the correlation between flow, cavitation, and wear to achieve joint diagnosis. Therefore, a non-invasive, multi-state intelligent diagnostic technology with a defined data processing link, capable of online deployment, and adaptable to complex and varying operating conditions is needed. Summary of the Invention

[0006] To address the problems of existing technologies, such as high dependence on external sensors, difficulty in extracting weak q-axis current features, incomplete single-domain feature representation, and insufficient diagnostic stability under multi-state coupling conditions, this invention provides a multi-state intelligent diagnostic method and system for centrifugal pumps under varying operating conditions based on the fusion of time-frequency features of motor q-axis current. By utilizing the current and synchronization angle information in the power supply circuit of the drive motor, synchronous diagnosis of flow rate, cavitation, and wear status under multiple operating conditions can be achieved without relying on additional sensors on the pump body or pipeline.

[0007] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0008] A multi-state intelligent diagnostic method for centrifugal pumps under varying operating conditions based on the fusion of time-frequency characteristics of motor q-axis current includes an offline model training stage and an online intelligent diagnostic stage, with the two stages connected by solidifying model parameters;

[0009] The offline model training phase includes the following steps:

[0010] The three-phase stator current, synchronous rotating coordinate system electrical angle and corresponding status labels of the centrifugal pump are collected under multiple operating conditions. The status labels include flow condition label, cavitation stage label and mouth ring wear status label.

[0011] Clarke transform and Park transform are performed sequentially on the three-phase stator currents to extract the q-axis current. The q-axis current is then windowed and preprocessed to obtain standardized time-domain samples.

[0012] Multi-level discrete wavelet decomposition is performed on the standardized time-domain samples, and the inverse transform of the detail coefficients of each level is performed to reconstruct a multi-channel time-domain reconstruction array. Fourier transform is performed on the time-domain reconstruction array to obtain a multi-channel frequency-domain amplitude array.

[0013] A multi-task deep diagnostic model is constructed, with time-domain reconstruction array and frequency-domain amplitude array as dual-branch inputs. After feature extraction, channel recalibration, bidirectional time series modeling and attention feature focusing, a shared state feature vector is obtained. Based on the shared state feature vector, multi-task classification results of flow conditions, cavitation stage and mouth ring wear state are output.

[0014] A multi-task deep diagnostic model is trained based on the state labels, and the optimal model parameters obtained from the training are solidified.

[0015] The online intelligent diagnostic phase includes the following steps:

[0016] The three-phase stator current and synchronous rotating coordinate system electrical angle in the power supply circuit of the centrifugal pump drive motor are collected in real time. Clarke transformation and Park transformation are performed in sequence to extract the q-axis current. The q-axis current is then windowed and preprocessed to obtain online standardized time-domain samples.

[0017] Perform the same multi-level discrete wavelet decomposition and Fourier transform on the online standardized time-domain samples as in the offline model training phase to obtain the online time-domain reconstruction array and the online frequency-domain amplitude array;

[0018] The online time-domain reconstruction array and online frequency-domain amplitude array are input into the trained multi-task deep diagnostic model, and the fixed optimal model parameters are called for forward inference to output online diagnostic results of flow conditions, cavitation stage and mouth ring wear status.

[0019] Furthermore, the status label is calibrated as follows:

[0020] Flow rate condition label: at 0.4Q d Up to 1.4Q d Several flow rate operating points are set at intervals within the range, Q d This refers to the rated flow rate of the centrifugal pump.

[0021] Cavitation stage label: calibrated based on the head descent rate ΔH. The expression for the head descent rate ΔH is:

[0022] ,

[0023] Where: H0 is the non-cavitation reference head; H1 is the current head;

[0024] When ΔH < 1%, it is defined as the normal stage; when 1% ≤ ΔH < 3%, it is defined as the initial stage; when 3% ≤ ΔH < 5%, it is defined as the development stage; when 5% ≤ ΔH < 7%, it is defined as the moderate stage; when ΔH ≥ 7%, it is defined as the severe stage.

[0025] Wear condition label for the mouth ring: includes uniform wear condition and eccentric wear condition. The uniform wear condition includes several different circumferential uniform gap values; the eccentric wear condition includes several different non-uniform gap combinations.

[0026] Furthermore, multi-level discrete wavelet decomposition is performed on the standardized time-domain samples to obtain multiple sets of detail coefficients layer by layer; inverse discrete wavelet transform is performed on each set of detail coefficients to reconstruct a time-domain signal of the same length as the original window; all reconstructed time-domain signals are stacked layer by layer to form a multi-channel time-domain reconstruction array; Fourier transform is performed on each channel of the multi-channel time-domain reconstruction array and the amplitude is taken; the amplitude spectra of all channels are stacked to form a multi-channel frequency domain amplitude array.

[0027] Furthermore, the multi-task deep diagnostic model includes a two-branch one-dimensional convolutional feature extraction module, an adaptive channel recalibration module, a bidirectional temporal modeling module, a multi-head attention feature focusing module, and three task output heads;

[0028] The dual-branch one-dimensional convolutional feature extraction module includes a temporal CNN-1D branch and a frequency CNN-1D branch, which respectively receive the multi-channel temporal reconstruction array and the multi-channel frequency amplitude array. After each undergoes multi-layer one-dimensional convolution, batch normalization, non-linear activation and pooling, the sequence length is unified by adaptive pooling and concatenated in the channel dimension to obtain fused features.

[0029] The adaptive channel recalibration module performs global pooling of the fused features along the sequence dimension to obtain a channel description vector, generates a channel weight vector based on the channel description vector, and multiplies the channel weight vector with the fused features element by element to obtain the recalibrated features.

[0030] The bidirectional temporal modeling module performs temporal dependency modeling on the recalibrated features along both the forward and reverse directions and then splices and fuses them.

[0031] The multi-head attention feature focusing module performs multi-head scaling dot product attention calculation on the output of the bidirectional temporal modeling module, and obtains a shared state feature vector through global pooling.

[0032] The three task output heads respectively map the shared state feature vector into the probability of flow condition category, the probability of cavitation stage category, and the probability of mouth ring wear state category.

[0033] Furthermore, in the adaptive channel recalibration module, the channel description vector is sequentially mapped through a first fully connected layer, a ReLU activation function, a second fully connected layer, and a Sigmoid function to generate a channel weight vector.

[0034] Furthermore, the bidirectional temporal modeling module includes a forward long short-term memory network layer and a backward long short-term memory network layer; the forward long short-term memory network layer propagates forward along the time step, and the backward long short-term memory network layer propagates backward along the time step; the hidden states output in the two directions are concatenated at each time step, and then compressed through a linear mapping layer to obtain temporal features.

[0035] Furthermore, in the multi-head attention feature focusing module, the output of the bidirectional temporal modeling module is linearly mapped to obtain the query matrix, key matrix, and value matrix respectively; each attention head independently performs scaled dot product attention calculation based on the query matrix, key matrix, and value matrix; the outputs of each head are concatenated and fused by linear mapping, and then global pooling is performed along the time step dimension to obtain a shared state feature vector.

[0036] Furthermore, a weighted multi-task loss function is used when training the multi-task deep diagnostic model. The weighted multi-task loss function is obtained by weighted summation of the flow condition classification loss, the cavitation stage classification loss, and the mouth ring wear state classification loss. Among them, the flow condition classification loss and the mouth ring wear state classification loss adopt the standard cross-entropy loss; the cavitation stage classification loss adopts Focal Loss to reduce the loss contribution of easily classified samples.

[0037] A system for a multi-state intelligent diagnostic method for centrifugal pumps under varying operating conditions based on the fusion of time-frequency characteristics of motor q-axis current includes:

[0038] The current acquisition module is used to acquire the three-phase stator current in the power supply circuit of the centrifugal pump drive motor;

[0039] The synchronization information acquisition module is used to acquire the electrical angle of the synchronous rotating coordinate system that is synchronized with the three-phase stator current.

[0040] The coordinate transformation module, connected to the current acquisition module and the synchronization information acquisition module, is used to perform Clarke transformation and Park transformation on the three-phase stator current and output the q-axis current.

[0041] The signal processing module, connected to the coordinate transformation module, is used to perform window segmentation and preprocessing on the q-axis current and output standardized time-domain samples.

[0042] The time-frequency dual-domain feature construction module is connected to the signal processing module. It is used to perform multi-level discrete wavelet decomposition on the standardized time-domain samples, perform inverse transform reconstruction on the detail coefficients of each level to obtain a multi-channel time-domain reconstruction array, and perform Fourier transform on the multi-channel time-domain reconstruction array to obtain a multi-channel frequency-domain amplitude array.

[0043] The multi-task deep diagnostic model, connected to the time-frequency dual-domain feature construction module, includes a dual-branch one-dimensional convolutional feature extraction module, an adaptive channel recalibration module, a bidirectional long short-term memory network temporal modeling module, a multi-head attention feature focusing module, and three task output heads, which are used to output flow condition categories, cavitation stage categories, and mouth ring wear state categories based on the multi-channel time-domain reconstruction array and the multi-channel frequency-domain amplitude array.

[0044] The results output module is connected to the multi-task deep diagnostic model and is used to display, store, or send the diagnostic results.

[0045] The beneficial effects of this invention are as follows:

[0046] 1. The centrifugal pump multi-state intelligent diagnosis method based on the fusion of time-frequency characteristics of motor q-axis current described in this invention, through the connection architecture of fixing model parameters in the offline model training stage and reusing the above parameters in the online intelligent diagnosis stage, can synchronously output the flow condition, cavitation stage and wear state of the mouth ring under the variable operating conditions by relying only on the three-phase stator current in the power supply circuit of the drive motor and the electrical angle of the synchronous rotating coordinate system.

[0047] 2. The intelligent diagnostic method for centrifugal pumps under varying operating conditions based on the fusion of time-frequency characteristics of motor q-axis current described in this invention extracts the q-axis current as a non-invasive characterization signal by performing Clarke and Park transformations on the three-phase stator current. This eliminates the need to add pressure, flow, vibration, and acoustic sensors to the pump body or pipeline, thus reducing the complexity of installation and maintenance.

[0048] 3. The intelligent diagnostic method for centrifugal pumps under varying operating conditions based on the fusion of time-frequency characteristics of motor q-axis current described in this invention performs multi-level discrete wavelet decomposition on standardized time-domain samples, performs inverse transform on the detail coefficients of each level to reconstruct a multi-channel time-domain reconstruction array, and performs Fourier transform on the time-domain reconstruction array to obtain a multi-channel frequency-domain amplitude array. This method simultaneously preserves the effects of cavitation transient impact, load envelope, frequency band energy, and sideband modulation information.

[0049] 4. The intelligent diagnostic method for multi-state centrifugal pumps under varying operating conditions based on the fusion of time-frequency features of motor q-axis current described in this invention sets an adaptive channel recalibration module at the back end of the dual-branch one-dimensional convolution feature extraction module. The fused features are globally pooled along the sequence dimension to obtain the channel description vector. Based on the channel description vector, a channel weight vector is generated and multiplied element-wise with the fused features. This can adaptively reduce the interference of inverter harmonics, random pulsations, and irrelevant frequency band responses on shared features according to the samples.

[0050] 5. The intelligent diagnostic method for multi-state centrifugal pumps under varying operating conditions based on the fusion of time-frequency features of motor q-axis current described in this invention performs time-series dependency modeling and splicing fusion along the forward and reverse directions of time steps using a bidirectional long short-term memory network. Then, it focuses on the features of key time steps through a multi-head scaling dot product attention mechanism, thereby utilizing the contextual information in the state evolution process and enhancing the expression of time-sensitive features.

[0051] 6. The intelligent diagnostic method for centrifugal pumps under varying operating conditions based on the fusion of time-frequency characteristics of motor q-axis current described in this invention sets up three task output heads in a unified shared state feature space: flow condition, cavitation stage, and mouth ring wear condition. It also uses a weighted multi-task loss function for end-to-end training, utilizing common information between tasks while retaining task-specific information to adapt to multivariate coupled operating conditions.

[0052] 7. The intelligent diagnostic method for centrifugal pumps under varying operating conditions based on the fusion of time-frequency characteristics of motor q-axis current described in this invention calculates and solidifies the global mean and global standard deviation based on the full training data during the offline model training stage. During the online intelligent diagnostic stage, the solidified statistics are directly called to perform standardization processing on the real-time window data. This achieves consistency between the statistical distribution of online input data and the offline training distribution, thereby avoiding the problem of distribution shift.

[0053] 8. The intelligent diagnostic method for centrifugal pumps under varying operating conditions based on the time-frequency characteristics of motor q-axis current described in this invention performs exponential smoothing on the online diagnostic results of multiple consecutive windows, and executes alarms or control interlocks when the cavitation stage reaches a moderate or higher level, the wear state of the mouth ring reaches a preset alarm level, or the flow condition deviates from the allowable range, thereby reducing false alarms caused by occasional noise and improving diagnostic results. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings described below are some embodiments of the present invention. For those skilled in the art, it is obvious that other drawings can be obtained from these drawings without creative effort.

[0055] Figure 1 This is a flowchart of the intelligent diagnostic method for multi-state changes in centrifugal pumps based on the fusion of time-frequency characteristics of motor q-axis current, as described in this invention.

[0056] Figure 2 This is a schematic diagram of the three-phase current coordinate transformation and q-axis current extraction of the present invention;

[0057] Figure 3 This is a schematic diagram of the construction of the time-frequency dual-domain feature of q-axis current based on discrete wavelet transform according to the present invention;

[0058] Figure 4 This is a schematic diagram of the time-frequency fusion multi-task diagnostic network structure of the present invention;

[0059] Figure 5 This is a schematic diagram illustrating the model training, deployment, and online diagnostic process of this invention. Detailed Implementation

[0060] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0061] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "axial," "radial," "vertical," "horizontal," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0062] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0063] The centrifugal pump is driven by a drive motor. The drive unit and the drive motor are connected via a power supply line to provide electrical energy to the drive motor. A three-phase current acquisition module is installed in the power supply line between the drive unit and the drive motor, or it can be located in the internal current sampling circuit of the drive unit. The three-phase current acquisition module includes a current sensor, an anti-aliasing filter unit, an analog-to-digital converter unit, and a synchronous clock unit, used to synchronously acquire the three-phase stator current i. a i b i c The instantaneous value of the electrical angle θ of the synchronously rotating coordinate system. e It can be directly output by the drive unit, or obtained based on the feedback signal of the encoder or position sensor, or calculated by an angle estimator based on the motor voltage and current model.

[0064] like Figure 1 and Figure 5As shown, the multi-state intelligent diagnostic method for centrifugal pumps under varying operating conditions based on the time-frequency characteristics fusion of motor q-axis current, as described in this invention, includes an offline model training stage and an online intelligent diagnostic stage. The offline model training stage completes the construction, training, and parameter fixing of the diagnostic model based on historically collected data. The online intelligent diagnostic stage loads the model parameters fixed in the offline model training stage based on real-time collected motor current signals, performs forward inference, and outputs the diagnostic results. The two stages are connected through the fixed model parameters and a unified feature processing flow. Specifically, it includes the following steps:

[0065] I. Offline Model Training Phase: This phase involves collecting and calibrating training data, partitioning the sample set, and training the deep neural network in an offline environment. The goal is to obtain an optimal set of model parameters for use in the subsequent online diagnostic phase. Details are as follows:

[0066] S01: Multi-condition training data acquisition and label calibration;

[0067] Sample data used for offline training was acquired through actual testing on a test bench. The test bench was equipped with a centrifugal pump, drive motor, drive unit, three-phase current acquisition module, flow meter, outlet pressure sensor, and inlet pressure sensor. The three-phase current acquisition module was installed in the power supply line between the drive unit and the drive motor, and was used to synchronously acquire the three-phase stator current i. a i b i c The flow meter is installed on the outlet pipe of the centrifugal pump to measure the actual flow rate. The outlet pressure sensor and inlet pressure sensor are installed at the outlet flange and inlet flange of the centrifugal pump, respectively, to measure the pump's inlet and outlet pressures and calculate the head. The synchronous rotating coordinate system uses an electrical angle θ. e The sampling frequency is directly output from the drive unit. The sampling frequency is determined based on the rated power of the drive motor and the switching frequency of the inverter; in this embodiment, the sampling frequency is 8000Hz. A synchronous clock unit ensures consistent sampling times for the three-phase current, and the cutoff frequency of the anti-aliasing filter unit is set to 0.4 times the sampling frequency.

[0068] It should be noted that the flow meter and pressure sensor are only used to provide true labels for the samples during the offline training phase. Once the model has completed training and been deployed to the online operating environment, the input to the diagnostic system is only the three-phase stator current in the drive motor power supply circuit and the electrical angle of the synchronous rotating coordinate system, and no longer relies on any additional sensors on the pump body or pipeline.

[0069] Each training sample simultaneously has three types of labels: traffic condition label y Q cavitation stage label y C Wear status label for mouth ring y W The labeling methods for the three types of labels are as follows:

[0070] Flow rate label: Denote the rated flow rate of the centrifugal pump as Q. d At 0.4Q d Up to 1.4Q d Within the range of 0.1Q d Eleven flow rate test points were set at intervals. Each sample was assigned one of 11 flow rate labels based on the interval of the actual flow rate measured by the flow meter at the time of sample collection.

[0071] Cavitation Stage Label: The cavitation stage is calibrated based on the rate of descent ΔH of the current head H1 relative to the non-cavitation reference head H0. The expression for the rate of descent ΔH is:

[0072] ,

[0073] Where: H0 is the non-cavitation reference head, that is, the head when the inlet valve is fully open and no cavitation occurs under the same flow rate and wear conditions, in meters; H1 is the current head, in meters; ΔH is the head reduction rate, dimensionless, expressed as a percentage.

[0074] The calibration rules for the five cavitation stage levels based on the head descent rate ΔH are as follows:

[0075] When ΔH < 1%, it is calibrated as the normal stage (NOR);

[0076] When 1%≤ΔH<3%, it is defined as the early stage (EAR);

[0077] When 3%≤ΔH<5%, it is designated as the development stage (EVO).

[0078] When 5%≤ΔH<7%, it is defined as the moderate stage (MOD);

[0079] When ΔH ≥ 7%, it is calibrated as the severe stage (SEV).

[0080] The head drop rate thresholds in the above five-level classification are set based on the cavitation performance judgment standards commonly used in the centrifugal pump industry. In practical applications, they can be appropriately adjusted according to the anti-cavitation performance of specific pump types and the allowable head margin of the system. There is a coupling relationship between the cavitation stage classification and the flow condition classification: under the same cavitation degree, the absolute value of the head drop rate at different flow points may be different, but classifying each stage according to the relative proportion of ΔH can unify the calibration scale under different operating conditions.

[0081] Wear Labels for Wear Rings: Wear conditions of wear rings are calibrated by replacing them with wear rings of different clearance sizes. Uniform wear is simulated by installing wear rings with a uniformly widening circumferential clearance, characterizing the state of the inner surface of the wear ring after uniform circumferential wear. Eccentric wear is simulated by installing wear rings with a non-uniform clearance distribution, characterizing the non-uniform clearance state caused by installation eccentricity, localized wear, or rotor eccentricity. The eight types of wear ring condition labels include four types of uniform wear and four types of eccentric wear; the corresponding clearance values ​​for each type are shown in Table 1.

[0082] Table 1. Examples of Wear Condition Labels for O-rings

[0083] avg1 uniform 0.3 ecc1 eccentric 0.3 / 0.5 avg2 uniform 0.7 ecc2 eccentric 0.3 / 0.7 avg3 uniform 0.9 ecc3 eccentric 0.3 / 0.9 avg4 uniform 1.1 ecc4 eccentric 0.3 / 1.1

[0084] In Table 1, the clearance values ​​corresponding to the uniform wear state are the single-sided clearance values ​​in the radial direction of the wear ring; "0.3 / 0.5" corresponding to the eccentric wear state indicates an eccentric clearance distribution with a minimum clearance of 0.3 mm and a maximum clearance of 0.5 mm, and the other eccentric states are deduced by analogy.

[0085] After completing the calibration of the above three types of tags, the three-phase stator current data i of each sample are... a i b i c and its corresponding synchronous rotating coordinate system electrical angle θ e Perform Clarke and Park transforms to extract the q-axis current i. q .

[0086] S02: Training sample window segmentation and preprocessing, as detailed below:

[0087] The training set, validation set, and test set are divided into three parts according to a 7:2:1 ratio, using complete continuous acquisition sequences or complete operating condition groups as the basic unit. Subsequently, the continuous q-axis current sequence is windowed within each data subset. Each window contains N=2048 sampling points, corresponding to a time length of 0.256 seconds at a sampling frequency of 8000Hz. Adjacent windows have a 50% overlap, meaning that a window is output by sliding every 1024 sampling points.

[0088] By first dividing the data into subsets and then splitting the windows, we can avoid the same continuous collection sequence or adjacent windows being distributed across the training set, validation set, and test set, thus preventing data leakage.

[0089] For each window containing the q-axis current data x(n), n = 0, 1, ..., N−1, x(n) represents the i-th current within that window. q The sampled values, in units of A, are processed sequentially with outlier removal, detrending, and standardization.

[0090] Outlier removal employs an outlier detection method based on the 3-standard-deviation criterion: the mean μ of all sampled points within the window is calculated. x and standard deviation σ x This will satisfy |x(n)-μ x |>3σ x Sampling points identified as outliers are removed and filled using linear interpolation. The linear interpolation method involves taking the nearest valid sampling points before and after the outlier and calculating the filled value using a linear weighted average based on the time distance.

[0091] Detrending processing eliminates the slow drift component in the q-axis current by subtracting the linear trend of the data within the window.

[0092] The standardization process uses Z-score standardization, whose expression is:

[0093] ,

[0094] in: This is the value of the nth sample point after detrending, in units of A; μ is the standardized dimensionless value. global This is the global mean of the entire training data after detrending, in units of A; The global standard deviation of the entire training data after detrending is expressed in A. To prevent division by zero constant, take =10 -8 .

[0095] The preprocessed q-axis current data constitute time-domain samples of uniform length. Where N=2048, the above μ global and σ global During the offline training phase, the data is statistically analyzed and solidified based on the full training data, and then reused during the online intelligent diagnosis phase.

[0096] S03: Construction of time-frequency dual-domain features of training samples, specifically:

[0097] Multi-level discrete wavelet decomposition is performed on the standardized time-domain samples. The discrete wavelet decomposition decomposes the signal layer by layer into low-frequency approximate components and high-frequency detail components through a pair of orthogonal mirror filter banks.

[0098] like Figure 3 As shown, multi-level discrete wavelet decomposition is performed on the standardized time-domain samples. Discrete wavelet decomposition uses a pair of orthogonal mirror filter banks to decompose the signal layer by layer into low-frequency approximate components and high-frequency detail components. The approximation coefficients c of the j-th level decomposition are... j,k and detail coefficient d j,k Obtained recursively from the low-pass and high-pass analysis filters respectively:

[0099] ,

[0100] Where: j is the decomposition level number, j=1,2,…,J, J is the total decomposition level; k is the time position index of the coefficient of each level; It is a scaling function; c is a wavelet function; j,k The approximation coefficients for the j-th layer have the same dimensions as the original signal x(n); d j,k is the detail coefficient of the j-th layer, which has the same dimensions as the original signal.

[0101] In this embodiment, the db4 wavelet basis is used, the total number of decomposition layers J=8, the detail coefficients of layers D1 to D8 are retained, and the approximation coefficients of the 8th layer are discarded.

[0102] For each level of detail coefficient d j,k Perform inverse discrete wavelet transform on each of the (j=1,2,…,8) components individually to reconstruct the corresponding signals of the detail components in the original time domain. n = 0, 1, ..., N−1. The length of each reconstructed signal layer is the same as the original window length, which is 2048 sampling points. The 8 reconstructed signals are stacked sequentially to form a multi-channel temporal reconstruction array X with a dimension of 8×2048. time .

[0103] The physical meaning of the time-domain reconstruction array is as follows: the smaller-scale detail components (D1, D2) mainly reflect high-frequency transient impact components and electromagnetic noise; the intermediate-scale detail components (D3, D4, D5) mainly reflect broadband transient disturbances caused by cavitation bubble collapse and pressure pulsation in the flow channel; the larger-scale detail components (D6, D7, D8) mainly reflect the load envelope caused by flow rate changes, low-order modulation near the frequency conversion, and slow load fluctuations caused by mouth loop leakage.

[0104] Subsequently, the temporal reconstruction array X time Each channel in the process undergoes a 2048-point real-valued Fast Fourier Transform, and the amplitude (modulus) at each frequency point is taken, retaining values ​​from 0 to the Nyquist frequency f. s A total of 1025 amplitude points are collected within the 4000Hz range, and these are stacked sequentially to form a multi-channel frequency domain amplitude array X with a dimension of 8×1025. freq .

[0105] The frequency domain amplitude array highlights the energy concentration characteristics of each detailed component at different frequency scales: cavitation impact leads to energy enhancement in the mid-to-high frequency band (1000-2000Hz), asymmetric flow caused by mouth ring wear is mainly reflected in the sideband modulation energy change in the mid-to-low frequency band (31.25-250Hz), and flow deviation from the design point is mainly manifested as amplitude shift of the load envelope in the low frequency band (below 62.5Hz).

[0106] The wavelet basis type db4 and the number of decomposition layers J = 8 are fixed for reuse in the online intelligent diagnosis stage.

[0107] S04: Construction, training, and parameter fixing of multi-task deep diagnostic models, as detailed below:

[0108] S4.1 Model structure construction, such as Figure 4 As shown, the multi-task deep diagnostic model includes a two-branch one-dimensional convolutional feature extraction module, a dynamic soft mask channel recalibration module, a bidirectional long short-term memory network temporal modeling module, a multi-head attention feature focusing module, and three task output heads.

[0109] The dual-branch one-dimensional convolutional feature extraction module includes a temporal CNN-1D branch and a frequency-domain CNN-1D branch. The input to the temporal CNN-1D branch is... The input to the frequency domain CNN-1D branch is .

[0110] Each branch contains a first convolutional layer and a second convolutional layer. The first convolutional layer has a kernel size of 16, a stride of 1, and 32 output channels, used to extract the envelope variation and energy distribution trend of the signal within a wide receptive field. Following the first convolutional layer are a batch normalization layer, a ReLU nonlinear activation layer, and a one-dimensional maximum pooling layer with a pooling window size of 2 and a stride of 2. The second convolutional layer has a kernel size of 8, a stride of 1, and 64 output channels, used to extract local transient impulse patterns and harmonic modes within a smaller receptive field. Following the second convolutional layer are also a batch normalization layer, a ReLU nonlinear activation layer, and a one-dimensional maximum pooling layer with a pooling window size of 2 and a stride of 2.

[0111] After each branch undergoes its own convolution operation, the lengths of the feature sequences change. To concatenate the outputs of the two branches along the channel dimension, an adaptive average pooling layer is applied at the end of each branch to unify the output sequence length of the temporal branch and the frequency domain branch to 32 time steps. At this point, each branch outputs 64 channels with a sequence length of 32. The triggering condition and branching logic are as follows: if the output sequence lengths of the temporal CNN-1D branch and the frequency domain CNN-1D branch are inconsistent, adaptive average pooling is used to unify the sequence lengths of the two branches to the same value before concatenating along the channel dimension; if the output sequence lengths of the two branches are already consistent, concatenation is performed directly.

[0112] Output of the time-domain branch and the output of the frequency domain branch The features are spliced ​​along the channel dimension to obtain the fused features. The fusion features preserve the alignment of time-domain and frequency-domain information along the time axis along the sequence dimension (32 time steps), while simultaneously accommodating time-domain envelope information and frequency band energy information along the channel dimension (128 channels).

[0113] To address inter-channel interference issues caused by random pulsations, inverter switching noise, and irrelevant frequency band responses in industrial environments, a dynamic soft mask based on channel attention is employed to fuse the feature F. fused Perform channel recalibration:

[0114] right Global average pooling is performed along the sequence dimension (32 time steps) to obtain the channel description vector. Its c-th element is:

[0115] ,

[0116] Where: z c The mean of 32 time steps over the c-th channel, with dimensions equal to F. fused Consistent; F fused (c,t) represents the value of the c-th channel and the t-th time step in the fused feature.

[0117] The channel description vector z is sequentially input into the first fully connected layer, the ReLU activation function, the second fully connected layer, and the Sigmoid function, and the output is the channel weight vector. Its expression is:

[0118] ,

[0119] in: This is the weight matrix of the first fully connected layer. δ represents the corresponding bias term; δ is the ReLU activation function. This is the weight matrix of the second fully connected layer. For the corresponding bias term; The function is the sigmoid function, whose output range is (0,1); each element m of M c ∈(0,1) represents the soft mask coefficient of the c-th channel, which is dimensionless.

[0120] Combine the channel weight vector M with the fusion feature F fused Element-wise multiplication yields the recalibration features. :

[0121] ,

[0122] in: This represents the value of the c-th channel at the t-th time step after recalibration. The physical meaning of this element-wise multiplication is: when the soft mask coefficient m of channel c...c When the value is close to 1, the information in channel c is retained and transmitted to subsequent networks; when m... c When the value approaches 0, the c channel is suppressed. Because m c It is adaptively generated based on the channel description vector z of the current sample, which can dynamically adjust the importance weight of the channel for each sample, thereby effectively suppressing noise-dominated channels and enhancing state-sensitive channels.

[0123] The recalibrated feature F′ is input into a Bidirectional Long Short-Term Memory (BiLSTM) network. The BiLSTM consists of a forward LSTM layer and a backward LSTM layer. The forward LSTM propagates forward along time steps t=1,2,…,32, and the backward LSTM propagates backward along time steps t=32,31,…,1. Each direction outputs a 128-dimensional hidden state. At each time step, the forward and backward hidden states are concatenated to form a 256-dimensional temporal feature, which is then compressed to 128 dimensions through a linear mapping layer. The output of the BiLSTM is shown below. The fusion features retain the forward and backward temporal dependencies on the timeline. That is, forward propagation captures the impact of historical information on the current state, and backward propagation captures the correction of the current state by future information. The combination of the two ensures that the feature representation at each time step takes into account the contextual information.

[0124] An 8-head scaled dot product attention mechanism (Multi-Head Self-Attention) is applied to the output H of the BiLSTM. The basic formula for scaling dot product attention is:

[0125] ,

[0126] Where: Q, K, and V are the query matrix, key matrix, and value matrix, respectively, all obtained from H through linear mapping; d k Let d be the dimension of the key vector, used to scale the dot product result to control the gradient magnitude. k =128 / 8=16 (i.e., the feature dimension of each attention head). The eight attention heads each use independent linear mapping parameters, and each head independently performs scaled dot product attention calculations. The outputs of each head are concatenated along the channel dimension and then fused using linear mapping. The advantage of multi-head attention is that different attention heads can focus on different temporal patterns. For example, one head might focus on the moment of high-frequency transient impact, another head might focus on the slow changing trend of the low-frequency load envelope, and a third head might focus on the multi-time-step correlation of the state transition phase.

[0127] The output of the multi-head attention is globally averaged along the time step dimension to obtain a 128-dimensional shared state feature vector. The shared state feature vector is the unified input for subsequent multi-task classification. It is extracted from the time-frequency dual-domain fusion features through the combined effects of dynamic channel recalibration and bidirectional temporal modeling. It retains multi-scale time-frequency information and has noise robustness and temporal context awareness.

[0128] The three task output heads are the flow condition classification head, the cavitation stage classification head, and the ring wear state classification head. Each task output head is a single-layer fully connected classifier. The flow condition classification head maps the shared state feature vector to an 11-dimensional output, which, after Softmax activation, yields 11 types of flow condition probability distributions. The cavitation stage classification head maps to a 5-dimensional output, which, after Softmax activation, yields 5 types of cavitation stage probability distributions. The ring wear state classification head maps to an 8-dimensional output, which, after Softmax activation, yields 8 types of ring wear state probability distributions.

[0129] S4.2: Model training, as detailed below:

[0130] The model training employs a weighted multi-task loss function. The total loss function is obtained by weighted summation of the losses from the three tasks:

[0131] ,

[0132] Where: L Q The loss value for the flow condition classification head is dimensionless; L C L represents the classification head loss value during the cavitation stage, which is dimensionless. W The loss value for classifying the wear condition of the mouth ring is dimensionless; λ Q , λ C , λ W These are the weighting coefficients corresponding to the three losses.

[0133] Flow condition classification loss L Q Classification of wear condition and loss L of the ring W Standard cross-entropy loss is used:

[0134] ,

[0135] Where: B is the number of samples in the current training batch, dimensionless; C is the total number of categories for this task; y i,c Let be the true label of the i-th sample in class c, with a value of 0 or 1, and be dimensionless. Let be the dimensionless probability of the i-th sample in class c predicted by the model.

[0136] Cavitation stage classification loss L C Focal Loss is used:

[0137] ,

[0138] Where: p t α is the model's predicted probability of the true class, dimensionless; α is the class balance coefficient, dimensionless; γ is the focusing parameter, dimensionless.

[0139] In the embodiment, λ Q = 1.0, λ C = 2.5, λ W = 2.5, γ = 2.5, and α is set based on the reciprocal of the number of samples in each class in the training set. The optimizer used is AdamW, with an initial learning rate of 4 × 10⁻⁶. -4 The weight decay coefficient is 3×10 -2 The dropout rate was 0.3, and the training epochs were 200. The model saved its optimal parameters based on the weighted average of the accuracy of the three tasks on the validation set.

[0140] S4.3: Parameter Fixation

[0141] The trained multi-task deep diagnostic model was evaluated using a test set. The test set evaluation metrics included classification accuracy for each of the three tasks and the overall multi-task metric. After successful validation, the optimal model parameters were archived as an inference model file. The inference model file included network weights, bias terms, batch normalized statistics, and the mean μ archived during the preprocessing stage. global and standard deviation σ global Wavelet basis type db4 and decomposition layer number J=8.

[0142] II. Online Intelligent Diagnosis Phase

[0143] S05: Obtain the three-phase stator current signal in the power supply circuit of the centrifugal pump drive motor and the electrical angle of the synchronous rotating coordinate system synchronized with the three-phase stator current signal, as detailed below;

[0144] In the power supply circuit of drive motor 2, the three-phase stator current i is synchronously acquired by the three-phase current acquisition module 4. a i b i c The sampling frequency is determined based on the rated power of the drive motor 2 and the switching frequency of the inverter. In this embodiment, the sampling frequency is 8000Hz. The synchronous clock unit ensures that the sampling time of the three-phase current is consistent, and the cutoff frequency of the anti-aliasing filter unit is set to 0.4 times the sampling frequency to prevent high-frequency components from aliasing into the sampling signal.

[0145] Electrical angle θ of the synchronous rotating coordinate system synchronized with the three-phase stator current e The output is directly from drive unit 3. When drive unit 3 uses vector control, the control system obtains the synchronous rotating coordinate system electrical angle θ for Park transformation through rotor flux observation, slip frequency calculation, or other flux position estimation methods. eand the electrical angle θ of the synchronous rotating coordinate system e Provided to the coordinate transformation module.

[0146] When using an encoder or rotary transformer to obtain the rotor mechanical angle θm, the electrical angle θ of the synchronous rotating coordinate system is determined based on the rotor mechanical angle, number of pole pairs, and slip angle estimation results. e .

[0147] When the drive unit cannot directly provide the electrical angle θ of the synchronous rotating coordinate system e Furthermore, when no shaft end angular position sensor is configured, the electrical angle θ of the synchronous rotating coordinate system can be obtained based on the voltage and current signals of the drive motor, using a model reference adaptive system, sliding mode observer, or other flux linkage position estimation methods. e The above-mentioned methods of obtaining the angles are all conventional technical means in this field and do not constitute a limitation on the present invention. They can be flexibly selected according to the on-site hardware conditions.

[0148] S06: Perform Clarke transformation on the three-phase stator current, and then perform Park transformation based on the electrical angle of the synchronous rotating coordinate system to obtain the d-axis current and q-axis current in the synchronous rotating coordinate system. Use the q-axis current as a non-invasive characterization signal of the centrifugal pump state, as follows:

[0149] like Figure 2 As shown, the Clarke transform is first performed on the three-phase stator current. The Clarke transform converts the current components in the stationary three-phase coordinate system (abc coordinate system) into the α-axis and β-axis components in the two-phase stationary coordinate system (α-β coordinate system), and its expression is as follows:

[0150] ,

[0151] Where: i a i b i c This represents the instantaneous value of the three-phase stator current, expressed in amperes (A). and The current component is given in a two-phase stationary coordinate system, and the unit is A.

[0152] Subsequently, based on the electrical angle θ of the synchronous rotating coordinate system synchronized with the stator current... e Perform the Park transformation. The Park transformation converts the current components in the two-phase stationary coordinate system into the d-axis and q-axis current components in the two-phase rotating coordinate system (dq coordinate system). Its expression is:

[0153] ,

[0154] In the magnetic field-oriented coordinate system, the d-axis is aligned with the rotor flux linkage direction, and the d-axis current component i dIt is mainly related to the establishment of excitation flux; the q-axis is orthogonal to the d-axis, and the q-axis current component i q With the electromagnetic torque T of the motor e Between satisfy Where p is the extreme logarithm, The d-axis flux linkage is used because changes in the hydraulic load torque on the pump side directly cause responsive adjustments in the motor's electromagnetic torque. Therefore, the q-axis current component is more sensitive to steady-state deviations and transient fluctuations in the load torque. Thus, this invention selects the i-axis flux linkage. q (n) serves as the original input signal for subsequent centrifugal pump condition diagnosis, where n is the index of the discrete sampling point.

[0155] The electromechanical coupling of this invention is based on the following: the hydraulic power P of the centrifugal pump h =ρgQH, considering pump efficiency η and rotational speed ω m At that time, the pump load torque T L It can be represented as:

[0156]

[0157] Flow rate Q, head H, and efficiency η vary with flow conditions, cavitation level, and ring clearance, causing steady-state deviations and transient fluctuations in load torque. The mechanical motion equations of the motor are:

[0158]

[0159] Under the condition that the rotor flux linkage is oriented and the rotor flux linkage amplitude is approximately stable within a diagnostic window, the electromagnetic torque T e With q-axis current component i q Approximately proportional In actual motors, flux saturation and cross-coupling between the d and q axes also exist. Therefore, the q-axis current and load torque constitute a dynamic mapping involving nonlinear modulation. Flow rate changes mainly cause changes in the average load level and low-frequency modulation of the q-axis current; cavitation bubble formation and collapse mainly cause transient impacts and mid-to-high frequency energy enhancement; leakage and asymmetric flow caused by wear of the mouth ring mainly form mid-to-low frequency load modulation. Thus, the q-axis current can serve as a common non-intrusive information carrier for the above multiple states.

[0160] S07: The q-axis current is divided into continuous windows, and the q-axis current data in each window is preprocessed to obtain time-domain samples of uniform length, as follows:

[0161] For the extracted q-axis current i q(n) Perform continuous window segmentation. Each window contains N=2048 sampling points, corresponding to a time length of 0.256 seconds at a sampling frequency of 8000Hz. Adjacent windows are set with 50% overlap, that is, one window is output every 1024 sampling points to improve the output refresh rate of diagnostic results during online inference.

[0162] For the q-axis current data within each window, outlier removal, detrending, and standardization are performed sequentially. The outlier removal and detrending processes are consistent with step S02 in the offline model training phase. Standardization uses the same Z-score standardization as in step S02, where the mean μ... global and standard deviation σ global Directly call the global statistics solidified in step S02.

[0163] S08: Online time-frequency dual-domain feature construction, as detailed below:

[0164] Multi-level discrete wavelet decomposition is performed on the standardized time-domain samples. The wavelet basis type adopts db4, which is fixed in step S03 of the offline model training stage, and the number of decomposition levels adopts J = 8, which is fixed in the offline model training stage. The detail coefficients of the eight levels from D1 to D8 are retained, and the approximation coefficients of the eighth level are discarded.

[0165] Each detail coefficient is individually subjected to an inverse discrete wavelet transform to reconstruct the corresponding signal of that detail component in the original time domain. The eight reconstructed signals are stacked sequentially to form a multi-channel online time-domain reconstruction array with a dimension of 8×2048.

[0166] Perform a 2048-point real-number fast Fourier transform on each channel in the online time-domain reconstruction array, take the amplitude of each frequency point, retain 1025 amplitude points in the range from 0 to Nyquist frequency, and stack them in sequence to form a multi-channel online frequency domain amplitude array with a dimension of 8×1025.

[0167] S09: Multi-task status diagnosis and result output

[0168] The online temporal reconstruction array is input into the temporal CNN-1D branch of the multi-task deep diagnostic model, and the online frequency domain amplitude array is input into the frequency domain CNN-1D branch. Internally, the model sequentially performs bi-branch one-dimensional convolutional feature extraction, dynamic soft-mask channel recalibration, bidirectional long short-term memory network temporal modeling, and multi-head attention feature focusing. Finally, three task output heads output the probability distributions for traffic conditions, cavitation stages, and ring wear states, respectively. All network weights in the model utilize parameters from the inference model file fixed during the offline model training phase.

[0169] The output results of multiple consecutive windows are exponentially smoothed, and the smoothed class probabilities are used as the final diagnostic basis.

[0170] When the cavitation stage reaches a moderate or higher level, the wear condition of the mouth ring reaches the preset alarm level, or the flow rate deviates from the preset allowable range, the result output module performs audible and visual alarms, control interlocks, historical record archiving, or remote communication operations.

[0171] In a verification example, the intelligent diagnostic method for multi-state changes in centrifugal pumps based on the fusion of time-frequency characteristics of motor q-axis current, as described in this invention, achieved a verification accuracy of 99.7% for the flow rate task, 99.8% for the wear ring task, and approximately 97.0% for the cavitation task. The test accuracy, macro-average precision, macro-average recall, and macro-average F1 of the cavitation task were 97.02%, 97.15%, 96.94%, and 97.04%, respectively. These results demonstrate that the technical solution of this invention can effectively extract multi-state information from the q-axis current and do not constitute a limitation on the applicable pump type, data scale, or performance indicators of this invention.

[0172] In some embodiments, the sampling frequency is not limited to 8000Hz and can be adjusted according to the rated speed of the centrifugal pump, the switching frequency of the frequency converter, and the characteristic frequency band of the state to be diagnosed. It is generally recommended that the sampling frequency be no less than 20 times the fundamental frequency of the drive motor. The window length is not limited to 2048 points and can be adjusted according to the sampling frequency and the required frequency resolution. The wavelet basis is not limited to db4 and can use sym4, coif3, or other wavelet bases with good time-frequency localization characteristics. The number of decomposition levels is not limited to 8 levels and can be determined according to the sampling frequency and the lowest characteristic frequency of the state to be diagnosed. A general principle is to ensure that the center frequency of the smallest scale detail component is higher than the frequency converter switching frequency and the center frequency of the largest scale detail component is less than 0.5 times the rotational frequency.

[0173] In some embodiments, the kernel size of the one-dimensional convolutional branch is not limited to 16 and 8, and can be adjusted according to the sampling frequency and signal feature scale. The number of output channels of each branch is not limited to 64, and can be increased or decreased according to the number of model parameters and the computing power of edge devices. The hidden layer dimension and the number of multi-head attention heads of the bidirectional long short-term memory network can be adjusted according to actual needs.

[0174] In some embodiments, the flow condition output is not limited to a discrete classification form; a continuous flow regression head can be added after the shared state feature vector to achieve continuous flow value estimation. The cavitation stage label calibration is not limited to the head drop rate; it can be based on the effective net positive suction head (NPSH) or manual verification results. The wear condition label calibration is not limited to the gap size; it can be based on leakage measurement results or actual wear measured after maintenance and disassembly.

[0175] In some embodiments, the intelligent diagnostic method can run independently on an edge computing device, or it can be integrated with a programmable logic controller, a distributed control system, a device management platform, or a digital twin platform. For multiple centrifugal pumps of the same or different models, the model parameters can be updated and adapted through transfer learning, domain adaptation, or incremental learning, while the inputs during the online inference stage remain the motor current and the electrical angle of the synchronous rotating coordinate system.

[0176] Although the q-axis current is the only non-invasive information source, in some scenarios, if vibration or pressure sensors are already installed on-site, these signals can be used as auxiliary inputs and fused with the q-axis current characteristics. Adding auxiliary sensor signals does not constitute a necessary technical feature; the use of auxiliary sensor signals should be considered a supplementary implementation method based on the aforementioned intelligent diagnostic method.

[0177] The intelligent diagnostic system for multi-state operation of a centrifugal pump based on the time-frequency characteristics fusion of the q-axis current of a motor, as described in this invention, includes: a current acquisition module for acquiring the three-phase stator current in the power supply circuit of the centrifugal pump drive motor; a synchronization information acquisition module for acquiring the electrical angle of the synchronous rotating coordinate system synchronized with the three-phase stator current; a coordinate transformation module connected to the current acquisition module and the synchronization information acquisition module for performing Clarke and Park transforms on the three-phase stator current and outputting the q-axis current; a signal processing module connected to the coordinate transformation module for windowing and preprocessing the q-axis current and outputting standardized time-domain samples; and a time-frequency dual-domain feature construction module connected to the signal processing module for processing the standardized... The time-domain samples undergo multi-level discrete wavelet decomposition, and the detail coefficients of each level are reconstructed by inverse transformation to obtain a multi-channel time-domain reconstruction array. A Fourier transform is then performed on the multi-channel time-domain reconstruction array to obtain a multi-channel frequency-domain amplitude array. A multi-task deep diagnostic model, connected to the time-frequency dual-domain feature construction module, includes a bi-branch one-dimensional convolutional feature extraction module, an adaptive channel recalibration module, a bidirectional long short-term memory network temporal modeling module, a multi-head attention feature focusing module, and three task output heads. These output heads are used to output flow condition categories, cavitation stage categories, and mouth ring wear state categories based on the multi-channel time-domain reconstruction array and the multi-channel frequency-domain amplitude array. A result output module, connected to the multi-task deep diagnostic model, is used to display, store, or send the diagnostic results.

[0178] The current acquisition module, the synchronization information acquisition module, and the coordinate transformation module are disposed in at least one of the driving device, the data acquisition device, or the edge computing device; the time-frequency dual-domain feature construction module and the multi-task deep diagnostic model are implemented by the processor executing a computer program in the memory.

[0179] The result output module includes at least one of a data display unit, a data storage unit, an anomaly alarm unit, and a communication unit; the communication unit is used to transmit the diagnostic results to an industrial control system, an equipment management platform, or a remote operation and maintenance platform.

[0180] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0181] The detailed descriptions listed above are merely specific illustrations of feasible embodiments of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-state intelligent diagnostic method for centrifugal pumps under varying operating conditions based on the fusion of time-frequency characteristics of motor q-axis current, characterized in that, It includes an offline model training phase and an online intelligent diagnosis phase, with the two phases connected by solidifying model parameters; The offline model training phase includes the following steps: The three-phase stator current, synchronous rotating coordinate system electrical angle and corresponding status labels of the centrifugal pump are collected under multiple operating conditions. The status labels include flow condition label, cavitation stage label and mouth ring wear status label. Clarke transform and Park transform are performed sequentially on the three-phase stator currents to extract the q-axis current. The q-axis current is then windowed and preprocessed to obtain standardized time-domain samples. Multi-level discrete wavelet decomposition is performed on the standardized time-domain samples, and the inverse transform of the detail coefficients of each level is performed to reconstruct a multi-channel time-domain reconstruction array. Fourier transform is performed on the time-domain reconstruction array to obtain a multi-channel frequency-domain amplitude array. A multi-task deep diagnostic model is constructed, with time-domain reconstruction array and frequency-domain amplitude array as dual-branch inputs. After feature extraction, channel recalibration, bidirectional time series modeling and attention feature focusing, a shared state feature vector is obtained. Based on the shared state feature vector, multi-task classification results of flow conditions, cavitation stage and mouth ring wear state are output. A multi-task deep diagnostic model is trained based on the state labels, and the optimal model parameters obtained from the training are solidified. The online intelligent diagnostic phase includes the following steps: The three-phase stator current and synchronous rotating coordinate system electrical angle in the power supply circuit of the centrifugal pump drive motor are collected in real time. Clarke transformation and Park transformation are performed in sequence to extract the q-axis current. The q-axis current is then windowed and preprocessed to obtain online standardized time-domain samples. Perform the same multi-level discrete wavelet decomposition and Fourier transform on the online standardized time-domain samples as in the offline model training phase to obtain the online time-domain reconstruction array and the online frequency-domain amplitude array; The online time-domain reconstruction array and online frequency-domain amplitude array are input into the trained multi-task deep diagnostic model, and the fixed optimal model parameters are called for forward inference to output online diagnostic results of flow conditions, cavitation stage and mouth ring wear status.

2. The intelligent diagnostic method for multi-state changes in operating conditions of a centrifugal pump based on the fusion of time-frequency characteristics of motor q-axis current according to claim 1, characterized in that, The status label is identified as follows: Flow rate condition label: at 0.4Q d Up to 1.4Q d Several flow rate operating points are set at intervals within the range, Q d This refers to the rated flow rate of the centrifugal pump. Cavitation stage label: calibrated based on the head descent rate ΔH. The expression for the head descent rate ΔH is: ， Where: H0 is the non-cavitation reference head; H1 is the current head; When ΔH < 1%, it is defined as the normal stage; when 1% ≤ ΔH < 3%, it is defined as the initial stage; when 3% ≤ ΔH < 5%, it is defined as the development stage; when 5% ≤ ΔH < 7%, it is defined as the moderate stage; when ΔH ≥ 7%, it is defined as the severe stage. Wear condition label for the mouth ring: includes uniform wear condition and eccentric wear condition. The uniform wear condition includes several different circumferential uniform gap values; the eccentric wear condition includes several different non-uniform gap combinations.

3. The intelligent diagnostic method for multi-state changes in operating conditions of a centrifugal pump based on the fusion of time-frequency characteristics of the motor q-axis current according to claim 1, characterized in that, The standardized time-domain samples are subjected to multi-level discrete wavelet decomposition, and multiple sets of detail coefficients are obtained by decomposing layer by layer. Each set of detail coefficients is subjected to inverse discrete wavelet transform to reconstruct a time-domain signal of the same length as the original window. All reconstructed time-domain signals are stacked layer by layer to form a multi-channel time-domain reconstruction array. Fourier transform is performed on each channel of the multi-channel time-domain reconstruction array and the amplitude is taken. The amplitude spectra of all channels are stacked to form a multi-channel frequency domain amplitude array.

4. The intelligent diagnostic method for multi-state changes in operating conditions of a centrifugal pump based on the fusion of time-frequency characteristics of motor q-axis current according to claim 1, characterized in that, The multi-task deep diagnostic model includes a two-branch one-dimensional convolutional feature extraction module, an adaptive channel recalibration module, a bidirectional temporal modeling module, a multi-head attention feature focusing module, and three task output heads; The dual-branch one-dimensional convolutional feature extraction module includes a temporal CNN-1D branch and a frequency CNN-1D branch, which respectively receive the multi-channel temporal reconstruction array and the multi-channel frequency amplitude array. After each undergoes multi-layer one-dimensional convolution, batch normalization, non-linear activation and pooling, the sequence length is unified by adaptive pooling and concatenated in the channel dimension to obtain fused features. The adaptive channel recalibration module performs global pooling of the fused features along the sequence dimension to obtain a channel description vector, generates a channel weight vector based on the channel description vector, and multiplies the channel weight vector with the fused features element by element to obtain the recalibrated features. The bidirectional temporal modeling module performs temporal dependency modeling on the recalibrated features along both the forward and reverse directions and then splices and fuses them. The multi-head attention feature focusing module performs multi-head scaling dot product attention calculation on the output of the bidirectional temporal modeling module, and obtains a shared state feature vector through global pooling. The three task output heads respectively map the shared state feature vector into the probability of flow condition category, the probability of cavitation stage category, and the probability of mouth ring wear state category.

5. The intelligent diagnostic method for multi-state changes in operating conditions of a centrifugal pump based on the fusion of time-frequency characteristics of the motor q-axis current according to claim 4, characterized in that, In the adaptive channel recalibration module, the channel description vector is sequentially mapped through the first fully connected layer, the ReLU activation function, the second fully connected layer, and the Sigmoid function to generate the channel weight vector.

6. The intelligent diagnostic method for multi-state changes in operating conditions of a centrifugal pump based on the fusion of time-frequency characteristics of the motor q-axis current according to claim 4, characterized in that, The bidirectional temporal modeling module includes a forward long short-term memory (LSTM) network layer and a backward LSM network layer. The forward LSM network layer propagates forward along the time step, and the backward LSM network layer propagates backward along the time step. The hidden states output from the two directions are concatenated at each time step and then compressed through a linear mapping layer to obtain temporal features.

7. The intelligent diagnostic method for multi-state changes in operating conditions of a centrifugal pump based on the fusion of time-frequency characteristics of the motor q-axis current according to claim 6, characterized in that, In the multi-head attention feature focusing module, the output of the bidirectional temporal modeling module is linearly mapped to obtain the query matrix, key matrix, and value matrix respectively. Each attention head independently performs scaled dot product attention calculation based on the query matrix, key matrix, and value matrix. The outputs of each head are concatenated and fused by linear mapping, and then global pooling is performed along the time step dimension to obtain a shared state feature vector.

8. The intelligent diagnostic method for multi-state changes in operating conditions of a centrifugal pump based on the fusion of time-frequency characteristics of the motor q-axis current according to claim 1, characterized in that, The multi-task deep diagnostic model is trained using a weighted multi-task loss function, which is obtained by weighted summation of flow condition classification loss, cavitation stage classification loss, and mouth ring wear state classification loss. The flow condition classification loss and mouth ring wear state classification loss use standard cross-entropy loss, while the cavitation stage classification loss uses Focal Loss to reduce the loss contribution of easily classified samples.

9. A system for intelligent diagnosis of centrifugal pumps under varying operating conditions and multiple states based on the fusion of time-frequency characteristics of motor q-axis current according to any one of claims 1-8, characterized in that, include: The current acquisition module is used to acquire the three-phase stator current in the power supply circuit of the centrifugal pump drive motor; The synchronization information acquisition module is used to acquire the electrical angle of the synchronous rotating coordinate system that is synchronized with the three-phase stator current. The coordinate transformation module, connected to the current acquisition module and the synchronization information acquisition module, is used to perform Clarke transformation and Park transformation on the three-phase stator current and output the q-axis current. The signal processing module, connected to the coordinate transformation module, is used to perform window segmentation and preprocessing on the q-axis current and output standardized time-domain samples. The time-frequency dual-domain feature construction module is connected to the signal processing module. It is used to perform multi-level discrete wavelet decomposition on the standardized time-domain samples, perform inverse transform reconstruction on the detail coefficients of each level to obtain a multi-channel time-domain reconstruction array, and perform Fourier transform on the multi-channel time-domain reconstruction array to obtain a multi-channel frequency-domain amplitude array. The multi-task deep diagnostic model, connected to the time-frequency dual-domain feature construction module, includes a dual-branch one-dimensional convolutional feature extraction module, an adaptive channel recalibration module, a bidirectional long short-term memory network temporal modeling module, a multi-head attention feature focusing module, and three task output heads, which are used to output flow condition categories, cavitation stage categories, and mouth ring wear state categories based on the multi-channel time-domain reconstruction array and the multi-channel frequency-domain amplitude array. The results output module is connected to the multi-task deep diagnostic model and is used to display, store, or send the diagnostic results.

Images